Eliminating the Auxiliary Clean-Label Anchor Required to Bootstrap Training under Imperfect Supervision

Requires meaningful augmentation pipelines that preserve label semantics; evaluation is limited to three regression datasets and does not address imbalanced label distributions (noted as future work)

Requires the label space to contain semantically opposing label pairs; without them the polarity vector is zero and divisiveness is trivially absent, making the method inapplicable

Relies on the SSL assumption that labeled and unlabeled data share class-conditional distributions, and its proxy validation set is resampled from a small labeled set, which may be unreliable when labeled data is extremely scarce

Relies on a pre-trained deep clustering model whose pseudo-labels must already have reasonable local consistency; cannot bootstrap from scratch

Performance on CN15k is notably poor for all methods; the authors attribute this to ConceptNet confidences being derived from source frequency with no real reliability distinction between high- and low-confidence triples, and call for better benchmark datasets

Requires access to a small set of input exemplars (recommended 64 per task) to compute feature covariance matrices, failing in strictly data-free deployment environments.

The exact adapted error level is infeasible to compute (Bernstein constants unknown); it is approximated by the classifier's probability mass on non-candidate labels, so the nominal coverage guarantee does not hold exactly in practice

Performance still falls well short of full supervision on most datasets (82.07 vs 90.11 average), since roughly half of the 'False' supervision conveys only negative information.

The claimed 70x human-labeling speedup assumes binary verification is a fixed unit time cost, which may not hold uniformly across datasets or annotators.

Requires a small labeled dataset (e.g., 100 samples/task) from the target tasks to compute the objective gradients, so it is not fully data-free

Relies on kernel density estimation, which scales poorly in high dimensions (many classifiers or classes)

Requires a task-specific validation set (~1,000 labeled samples), which can be expensive or infeasible to create for some tasks

Class-balancing strategy relies on the accuracy of pseudo-labels for recently annotated samples.

Requires labeled data and a trained GNN on the full primary dataset before subsampling, so it cannot be applied when labels are unavailable or the primary dataset is already too large/expensive to label

Relies on ERM-learned features, which may underperform in low-data or weak-label settings

Relies on an instance-dependent assumption p(ȳ|x,Y)=p(ȳ|x), which the authors admit is a tractable approximation and may fail when annotators cannot exclude labels without knowing the full relevant set

Requires warm-up training with standard cross-entropy on the noisy labels, which may still fit noise before robust training starts

Moving-average pseudo-label updates degrade performance substantially (e.g., ~21-point drop on F-MNIST-1 soft+adaptive), so the framework relies on careful per-dataset configuration; only basic pseudo-labeling techniques are studied

More computationally expensive than standard supervised learning, as acknowledged by the authors (requires fine-tuning multiple class-specific LLMs plus iterative SSL training)

Requires prior knowledge of the label distribution family (e.g., how many top entries to exclude) to design the variance loss; without such a prior, a range of λr values produce valid non-negative, sum-to-1 labels, so recovery is not guaranteed

Results come from a single train/validation split per dataset; verifying across multiple splits was computationally prohibitive (20,000+ GPU hours already spent).

The model remains poorly calibrated: calibration is measured dataset-wide rather than sample-wise, which may give a false sense of confidence reweighting

Method requires label prototype and per-label angle statistics to be updated every epoch, adding per-label bookkeeping for large label spaces (only tested up to 85 classes)

Requires enough samples per discretized label fragment for reliable expert training and neighborhood agreement, so sparse or unevenly distributed labels may fragment poorly

Assumes labeled and unlabeled samples share the same category space; open-set settings are explicitly left to future work, so the estimated label encodings are invalid if unknown classes exist in unlabeled data.

Requires extra weak annotations (point locations plus categories) for the otherwise unannotated images, so it is not applicable when no annotation at all is available for unlabeled data.

Treats outliers as unstructured perturbations, so it cannot learn the instance-dependent confusion function E(x_n), which could otherwise enable pre-annotation detection of difficult samples

It operates on shallow multi-label features that 'may be less informative'; no deep feature representation is used, limiting recognition quality

Fixing the encoder means it cannot adapt to newly acquired labels, which the authors concede might sometimes be suboptimal, though they only gesture at semi-supervised representation learning as a remedy.

Assumes all users are homogeneous (same skill), which is unrealistic since users have distinct classification accuracies; this avoids combinatorial explosion but limits real-world applicability

Relies on the assumption that integrated labels equal true labels after instance weighting; if label integration is poor, the bias-variance trade-off can amplify wrong labels

Requires access to the original training set at edit time to sample regularization batches, which may not be feasible in all deployment settings.

Requires CLIP feature extraction quality and Gaussian-kernel median bandwidth over the whole dataset (O(mn) plus kernel computation), which may scale poorly to very large unlabeled pools

Relies on synthetic partial-label generation (flipping negative labels with probability q) on small image benchmarks; no experiments on real-world partial-label datasets or large-scale data

Restricted to binary classification: in the multi-class setting, when all annotators are wrong on a disagreement sample, LRD is forced to select one of the incorrect labels, degrading performance.

Assumes the chosen reference input is irrelevant to class labels; this can fail when classification depends on color, requiring an out-of-distribution 'non-image' input as a workaround.

Requires multiple annotations per image (at least several per image) and iterative retraining; single-annotation or one-pass settings are not addressed

Assumes the number of label classes is known a priori and cannot handle growing label spaces in evolving environments.

Does not model label correlations or instance similarities, so it cannot exploit label co-occurrence structure

Not tested on related weak-supervision settings such as partial labels (incomplete annotation), single-positive labels, or general noisy labels where every label may be corrupted

Cross-dataset label inconsistency still causes notable accuracy drops (2.71% on RAF-DB), which the authors acknowledge remains an open challenge

Requires pre-collected environment labels, which add annotation cost; without them generalization is claimed theoretically impossible, so the method cannot run on datasets lacking environment information

Requires class-aware sampling of enough labeled minority samples to form balanced batches, which may fail under extreme scarcity (e.g., 3 labeled tail samples in some STL-10 settings)

Uses pre-defined fixed confidence thresholds (τ_p=0.95, τ_q=0.5) shared across all classes, which the authors note may be less flexible in more complex tasks

Applying A' to all samples destabilizes training and reduces performance (92.98% vs 92.50% baseline on CIFAR-10, worse on STL-10: 83.19% vs 85.27%), because it corrupts semantic information of non-naive samples.

Requires an auxiliary clean dataset (m ≪ n) and assumes the noise model satisfies that for each class the most probable corrupted label matches the true label; performance on other noise structures is not tested

Only evaluated on two benchmark datasets with modest identity counts; pseudo-label accuracy converges below full supervision (Fig. 4)

Adds per-epoch evaluation overhead: for each class, T(l+1) augmented samples must be classified to update LoL scores, and performance degrades if γ, p_aug, or T are set to make augmentation too strong or too weak

Class-imbalanced noisy-label settings are not considered (explicitly stated by the authors)

Performance on single positive settings still falls well short of full-label training (e.g., 72.89 vs 76.78 mAP on COCO), so the approach only partially bridges the gap

Assumes the available sparse labels cover all semantic classes present in the data; fails when classes are missing or novel

Not evaluated under class-imbalanced conditions or instance-dependent candidate-label generation

Requires reliable empirical estimates of class priors, which may degrade if tail classes have extremely sparse observations.

Requires a perfectly predefined and accurate label hierarchy (tree or DAG) to be effective.

Requires static knowledge of class frequencies from the training set, making it incompatible with online learning or streaming data.

Assumes complementary labels are chosen uniformly at random among non-true classes, failing in settings with annotator class-selection bias.